A joint project of the Graduate School, Peabody College, and the Jean & Alexander Heard Library

Title page for ETD etd-11192018-162105


Type of Document Master's Thesis
Author Koola, Jejo David
Author's Email Address jkoola@ucsd.edu
URN etd-11192018-162105
Title Risk Prediction Models and Visualizations for Hepatorenal Syndrome
Degree Master of Science
Department Biomedical Informatics
Advisory Committee
Advisor Name Title
Michael E. Matheny Committee Chair
Brad Malin Committee Member
Daniel Fabbri Committee Member
Keywords
  • cirrhosis
  • hepatorenal syndrome
  • risk prediction
  • machine learning
  • natural language processing
  • dimension reduction
  • variable selection
  • information visualization
Date of Defense 2016-08-26
Availability unrestricted
Abstract
Cirrhosis, a late stage of chronic liver damage where scarring replaces hepatic tissue, carries significant morbidity and mortality. The prevalence is estimated between 400,000 and 3,000,000 persons in the United States, and the disease causes 44,000 deaths annually. Cirrhosis impacts the health care system broadly because of the breadth and severity of end-stage liver disease complications. Hepatorenal syndrome (HRS) is a particularly challenging complication of end-stage cirrhosis, and represents an archetype of multi-organ failure. In the increasingly complicated, data-driven clinical environment informatics solutions may help improve care for patients with HRS. This thesis first strives to build an Electronic Health Record phenotyping model for HRS using Natural Language Processing and sophisticated machine learning techniques. Phenotyping has played an increasingly important part in observational cohort studies by allowing precise selection of cases and controls for further analysis. This is one of the first efforts to phenotype an acute kidney injury etiology, a condition that effects up to 2% of hospitalized patients. The penalized logistic regression achieved the best performance with an AUC of 0.93 (95% CI: 0.92-0.93). Subsequently, we develop a risk prediction model to identify patients at high likelihood of developing HRS during hospitalization based on information available within twenty-four hours of coming to the emergency room. Our model achieved good discrimination (AUC of 0.84). In addition to prediction performance, our study highlighted potentially new risk factors including Mean Corpuscular Hemoglobin Concentration and paracentesis. Finally, we designed interactive information visualization tools to help both researchers and clinicians better understand model performance. Using this end-to-end pipeline we believe we can improve the care of patients with HRS.
Files
  Filename       Size       Approximate Download Time (Hours:Minutes:Seconds) 
 
 28.8 Modem   56K Modem   ISDN (64 Kb)   ISDN (128 Kb)   Higher-speed Access 
  VanderbiltDissertationTemplate_final_submitted.pdf 5.02 Mb 00:23:13 00:11:56 00:10:27 00:05:13 00:00:26

Browse All Available ETDs by ( Author | Department )

If you have more questions or technical problems, please Contact LITS.